Introduction

Hi, I'm Matt, and welcome to the Metropolis blog.

Metropolis is an online dating app that uses machine learning to predict whether two people will be attracted to each other. The purpose of this blog is to talk about the app, and online dating in general.

But first, a bit about myself.

Until recently I was a software engineer at Google, working on Google Maps in the Sydney office. Prior to that I've been a hedge fund quant, a codebreaker with the Australian defence forces, and was a co-founder of MessageMedia, which recently sold for USD 1.3 billion (I was a contractor, so no equity unfortunately!)

I also have an MBA from Melbourne Business School and a PhD in Geospatial Science. LinkedIn profile here.

The origin story

While I was at Google, Sundar announced that we would become an AI-first company (pivoting from being mobile-first). All of us engineers were expected to become proficient in machine learning.

After completing the training I found there wasn't much immediate need for AI in Maps (although that has changed recently). So to make use of my skills I started playing with convolutional neural networks in my spare time, and managed to build a model that can measure the attractiveness of a human face.

It worked better than anything previously reported, and I thought maybe I could build a company around it. So I left Google and here we are.

I took this mirror selfie (left) to see if the model worked for mirror-image faces. It does.

It turns out a typical face is symmetric enough that taking the mirror image doesn't make any difference to its attractiveness.

Comments

Post a Comment

Popular posts from this blog

Are average faces more attractive?

Image
  In 1878, when Francis Galton discovered how to overlay photographic images, he observed that an averaged  face is more attractive than an average  face. In other words, averaging together lots of faces will result in a face that's more attractive than the average (or median) face in the original set — although not as attractive as the most attractive one. The effect was re-discovered in  1990 , and, ever since, researchers have speculated as to why. However, a problem with these studies is the way the faces are "averaged". Typically, it means calculating the average pixel values of the faces. But unless you first align the faces, all you get is a pinkish-brown blur. So researchers first resize and align the images so the eyes and mouths line up. Highlight the features, hide the flaws The result is faces where the eyes and mouths are clearly defined, and everything else is somewhat blurry. Coincidentally, this is what women do when they put on makeup. Or early movie dir
Read more

The geography of adult tech

Image
If you exclude Chinese firms, the world's biggest tech companies are all based in Silicon Valley (San Francisco Bay Area) or Seattle. The trillion-dollar club are Apple , Bay Area Amazon , Seattle Microsoft , Seattle Alphabet (Google et al), Bay Area Meta (Facebook et al), Bay Area And there are plenty of other Bay Area firms in the next tier, such as Twitter, Uber, and Netflix. Now let's look at the world's biggest adult-related tech firms. Match Group  (Tinder, Hinge, et al), Dallas, Texas Bumble , Austin, Texas MindGeek  (Pornhub, YouPorn, et al), Montreal, Canada OnlyFans , London, UK Tumblr  (pre-2018 family-friendly reboot), New York These are all multi-billion-dollar businesses, but notice that none of them are located on the US west coast. OK, technically Bay Area firms like Twitter and Reddit tolerate adult content, but it's a small part of their traffic, and they're rather coy about it. It's not like a Silicon Valley venture capitalist to pass up a
Read more

Race and attraction

Image
For many years the online dating site OkCupid produced the OkTrends blog, where they analysed their data to explore people's dating habits. In 2009 they looked at racial preferences — the original blog has disappeared, but a copy of the article can be found here . The article analysed the data through the lens of American racial politics, and, perhaps unsurprisingly, concluded that their users are racist. However, there are a few problems with their analysis. First, they measured preferences by looking at how frequently people responded to initial messages. However, this measures not just the preferences of the recipient, but the judgement of the sender. For example, a sender who mostly contacts attractive people will almost certainly get a lower response rate than one who contacts unattractive ones. Does judgement vary by race? We have no way of knowing. Second, much of their analysis is based on men responding to women's messages. But in online dating — as in the real world —
Read more